Changeset 096a174

Timestamp:

Jan 7, 2019, 10:55:27 PM (5 years ago)

Author:

Paul Brossier <piem@piem.org>

Branches:

feature/cnn_org, feature/crepe_org

Children:

147afba

Parents:

855947d

Message:

[fmat] use gemv in fmat_vecmul

This is possible since we use contiguous memory from 1fcd3928f.

Location:

src

Files:

• aubio_priv.h (modified) (2 diffs)
• fmat.c (modified) (3 diffs)

src/aubio_priv.h

@@ -132 +132 @@
 #define aubio_cblas_swap      cblas_sswap
 #define aubio_cblas_dot       cblas_sdot
+#define aubio_cblas__gemv     cblas_sgemv
 #else /* HAVE_AUBIO_DOUBLE */
 #ifdef HAVE_ATLAS
@@ -139 +140 @@
 #define aubio_cblas_swap      cblas_dswap
 #define aubio_cblas_dot       cblas_ddot
+#define aubio_cblas__gemv     cblas_dgemv
 #endif /* HAVE_AUBIO_DOUBLE */
 #endif /* HAVE_BLAS */

src/fmat.c

@@ -154 +154 @@
 
 void fmat_vecmul(const fmat_t *s, const fvec_t *scale, fvec_t *output) {
-  uint_t k;
-#if 0
-  assert(s->height == output->length);
-  assert(s->length == scale->length);
-#endif
 #if !defined(HAVE_ACCELERATE) && !defined(HAVE_BLAS)
-  uint_t j;
+  uint_t j, k;
+  AUBIO_ASSERT(s->height == output->length);
+  AUBIO_ASSERT(s->length == scale->length);
   fvec_zeros(output);
   for (j = 0; j < s->length; j++) {
     for (k = 0; k < s->height; k++) {
-      output->data[k] += scale->data[j]
-          * s->data[k][j];
+      output->data[k] += scale->data[j] * s->data[k][j];
     }
   }
 #elif defined(HAVE_BLAS)
+#if 0
   for (k = 0; k < s->height; k++) {
     output->data[k] = aubio_cblas_dot( s->length, scale->data, 1, s->data[k], 1);
   }
+#else
+  aubio_cblas__gemv(CblasColMajor, CblasTrans,
+      s->length, s->height, 1.,
+      s->data[0], s->length,
+      scale->data, 1, 0.,
+      output->data, 1);
+#endif
 #elif defined(HAVE_ACCELERATE)
 #if 0
@@ -177 +181 @@
   vDSP_mmul (s->data[0], 1, scale->data, 1, output->data, 1, s->height, 1, s->length);
 #else
+  uint_t k;
   for (k = 0; k < s->height; k++) {
     aubio_vDSP_dotpr( scale->data, 1, s->data[k], 1, &(output->data[k]), s->length);
@@ -182 +187 @@
 #endif
 #endif
+  }
+#endif
+#endif
 }